emperor-mew/voidly-zero-shot-transfer-v1

Voidly Atlas Zero-Shot Country Transfer v1

Version: v1 | License: CC BY 4.0

Meta-feature regressor predicting per-country LOCO F1 from country meta-features (measurement density, source diversity, regime cluster, neighbour topology). Used to forecast forecastability — i.e. predict which under-labeled countries will benefit most from new labels.

Eval

Metric	Value
r_squared	0.4900

Honest caveats

• R² 0.49 — moderate. Useful for targeted labeling priorities, not for production scoring.
• Trained on 131 LOCO points — small. Bootstrap CI is wide.

Citation

@misc{voidly_voidly_zero_shot_transfer_v1,
  title  = {Voidly Atlas: voidly-zero-shot-transfer-v1 (v1)},
  author = {Voidly},
  year   = {2026},
  url    = {https://huggingface.co/emperor-mew/voidly-zero-shot-transfer-v1},
  note   = {Open censorship-research ML stack. CC BY 4.0.}
}